lyraDiff IP-Adapter Weights Bundle

A ready-to-use weights bundle of IP-Adapter-Plus (SDXL, ViT-H) checkpoints and their CLIP image encoders, packaged for out-of-the-box use with the lyraDiff acceleration engine.

lyraDiff is a recompilation-free inference engine for Diffusion and DiT models that delivers state-of-the-art speed (up to ~6x) and memory savings via quantization, fused GEMM kernels, and flash attention while preserving pixel-level image consistency. This repo collects the IP-Adapter weights in the directory layout lyraDiff expects, so you can drop them in without re-downloading or re-organizing from multiple upstream sources.

Contents

ip_ckpt/lyra_tran/
├── models/image_encoder/              # CLIP image encoder (ViT-H)
└── sdxl_models/
    ├── image_encoder/                 # SDXL CLIP image encoder
    ├── ip-adapter-plus_sdxl_vit-h.bin # IP-Adapter-Plus weights
    └── ip-adapter-plus_sdxl_vit-h/    # unpacked cross-attention IP weights

Credits

• IP-Adapter — Tencent AI Lab
• lyraDiff — TMElyralab

This repository redistributes the above weights for convenience; all rights and licenses belong to their respective upstream authors.